Botir's AI Hardware Co.

Built on NVIDIA's Blackwell architecture with 21,760 CUDA cores. This is the card gamers and hobbyists fight over — and the honest truth is that for the big open models in this store, one of these is a starting point for experiments, not a production machine.

GPU memory: 32 GB GDDR7

The model must fit here, not in regular RAM. GDDR7 is fast graphics memory, but 32 GB holds only a small model — the big ones in our advisor need hundreds of gigabytes.

Power draw: 575 W

About half of what an average home draws around the clock (1,200 W). Your wall socket can handle one of these; your electricity bill will notice it.

Interconnect: PCIe 5.0 only — no NVLink

This card cannot be wired directly to another card. Two 5090s in one PC talk through the motherboard, which is far too slow to act as one big GPU.

575 W — about 48% of what an average home draws around the clock, and a day of running it uses about 15% of a 90 kWh electric-car battery.

NVIDIA's professional workstation card. Same Blackwell chip family as the 5090, but with 96 GB of error-correcting memory and drivers certified for professional work. One of these runs mid-size models on its own; the big models still need more.

GPU memory: 96 GB GDDR7 ECC

ECC means error-correcting: the memory catches and fixes the rare bit flips that would crash a week-long job. 96 GB fits models up to roughly 80 billion parameters.

Power draw: 600 W

Half a home's average draw. Runs on a normal workstation power supply.

Interconnect: PCIe 5.0 only — no NVLink

Like the 5090, these cards cannot be fused into one big GPU. A pile of them shares a motherboard, not memory.

600 W — about 50% of what an average home draws around the clock, and a day of running it uses about 16% of a 90 kWh electric-car battery.

A datacenter GPU on a PCIe card, made to be installed in server chassis in groups of two or four joined by NVLink bridges. This is the cheapest honest way into the hundreds-of-gigabytes club: three of them hold 423 GB.

GPU memory: 141 GB HBM3e

HBM is high-bandwidth memory, stacked right next to the chip. It moves data several times faster than the GDDR7 on desktop cards — and feeding the chip fast enough is the whole game.

Power draw: 600 W

Same as one workstation card — but with 1.5× the memory of a PRO 6000 and much faster memory at that.

Interconnect: NVLink bridge, 900 GB/s

Up to four H200 NVLs in one server are bridged so they share each other's memory at 900 GB/s. That is what lets three cards act as one 423 GB GPU.

600 W — about 50% of what an average home draws around the clock, and a day of running it uses about 16% of a 90 kWh electric-car battery.

NVIDIA's flagship single server: 8 Blackwell Ultra GPUs fused by NVLink into 2,304 GB of shared memory. Every open model on today's leaderboard fits in this box with room to spare. This is what 'serious capacity' means in one purchase order.

GPU memory: 2,304 GB HBM3e (8 × 288 GB)

More than two terabytes of GPU memory acting as one pool. The largest open model in our advisor needs 904.8 GB — this box holds it twice over.

Power draw: ~14,000 W

About 12 homes' worth of constant draw. It needs datacenter power and cooling — this does not plug into a wall.

Interconnect: 5th-gen NVLink, 1.8 TB/s per GPU

Every GPU talks to every other at 1.8 TB/s. This is what makes eight chips behave as one giant one — and what a pile of desktop cards can never do.

14,000 W — as much power as about 12 average homes, and a day of running it drains about 3.7 full 90 kWh electric-car batteries.

A full rack-scale system: 72 Blackwell Ultra GPUs and 36 Grace CPUs, liquid-cooled, wired by NVLink into a single 20.7 TB machine. This is what AI labs and clouds buy by the row. You buy one when you serve a big model to a lot of users at once.

GPU memory: 20,736 GB HBM3e (72 × 288 GB)

Nearly 21 terabytes of GPU memory in one machine — enough to hold the biggest open model about 23 times over, which is capacity for serving many users, not just fitting the model.

Power draw: ~135,000 W

As much power as about 112 homes, around the clock. One rack, a whole neighborhood's electricity.

Interconnect: NVLink across all 72 GPUs, 130 TB/s total

The entire rack is one NVLink domain — all 72 GPUs share memory as one machine. This is real clustering, done at the factory.

135,000 W — as much power as about 112 average homes, and a day of running it drains about 36.0 full 90 kWh electric-car batteries.